TDA4AH-Q1: When we test our model in SDK 9.2, we find that custom layer cannot remove the pad of the model output

Hi 

Why are you trying to remove the padding in model inference process?

We have provided an example to get the data without padding: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_io_tensors_format.md#input-and-output-tensor-formats-for-tidl-rt-inference 

Regards,

Adam

I want to fist understand the actual problem statement, please confirm if this is correct understanding

• You have a caffe network which naturally outputs some tensor shape which is having padded data.
• You want to remove the padded data towards the  end and you are using a custom layer for this

If yes then, you have several options ( in order of below preference)

1. Add crop layer in your caffe network so naturally it gives correct output

2. Have this processing on ARM or outside network even on C7x (without adding as a custom layer) - because it is anyway at the end

3. Add custom layer  - this is what you have tried

Can you first check the option 1 which is more natural and cleaner solution

Hi

Your understand is right. We will try to solve this problem by adding crop layer in the caffe network as you suggested

But we have the following doubts:

1.It is normal in SDK8.6 when we use the same method to remove pad in the same model, so it is a version compatibility issue. We hope that it can be resolved

2.We still don't know what caused the error this time. So we are not sure if we will encounter the same issue when adding a custom layer in the future.

Best regards

Guang Ren

Hi

We have tried to add crop layer in the caffe network as you suggested, but the pad still exists in the output of ebd_relu layer.

this is our method of modification:

/cfs-file/__key/communityserver-discussions-components-files/791/new_5F00_deploy.zip

Hi

By the way, may I ask What is the principle of crop layer removing pad? Is there any readily available material that I can refer to?

Best regards,

Guang Ren

Hi Guang,

Since this is, in the end, anyhow, I prefer to do a post-process to massage the output tensors.  You will have full control of how it is done.  The only downside is that it is done on the ARM core; it may be a bit slower, but since it is the last layer, I do not think it is a big difference. 

Regards,

Chris

Hi,

In our project, since the ARM core needs to handle other tasks and there are no remaining resources on the ARM core, we have to perform the operations on the dsp core. We hope you can help us resolve the issue of removing the pad by crop layer.

Regards,

Guang Ren

Hi Guang,

I do not think it will make a difference.  The Crop layer is not supported in TIDL and will be run on the ARM core anyhow.   Here is the ONNX and TIDL supported layers documentation.  Again, I would recommend doing this post-processing without a crop layer.

ONNX:

https://onnx.ai/onnx/operators/onnx__CenterCropPad.html 

TIDL:

https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/supported_ops_rts_versions.md

Regards,

Chris

Hi GUANG,

We are potentially looking for a solution which can remove the channel padding for the final output so you don't need a custom layer and also any post processing in ARM to remove the data between channels

This will require a change in our tool chain and we are in process of validating with change

Thanks

With Regards

Pramod